1 - Write out what you currently expect your AI agent to handle, in full.

A Perfect Prompt: Act as a role design consultant. Here is what I currently ask my AI agent or assistant to do: [DESCRIBE EVERYTHING]. Convert this into a formal job description, then tell me honestly whether a competent human could reliably perform this role as written. Ask me any questions you have.

[DESCRIPTION]. Break it into the distinct skills required, mark each as routine or judgment-heavy, and tell me how many separate roles this would be in a company with real headcount. Ask me any questions you have.

3 - Cut everything that requires judgment about your specific customers.

A Perfect Prompt: Act as an automation boundary specialist. Here is a list of responsibilities: [PASTE LIST]. Separate them into work that follows repeatable rules versus work that requires knowing a specific customer, relationship, or context. Explain the cost of getting each wrong. Ask me any questions you have.

4 - Cut everything that requires the agent to decide when to act.

A Perfect Prompt: Act as an AI reliability engineer. Here are the remaining responsibilities for my agent: [PASTE LIST]. Identify which ones require the agent to decide when to trigger versus which are triggered by a clear external event. Explain why the first category is where agents drift. Ask me any questions you have.

5 - What is left is your actual agent. Build only that.

A Perfect Prompt: Act as an AI workflow designer who builds narrow, reliable systems. Here is the single job I want my agent to do: [ONE SENTENCE JOB]. Design the complete workflow including trigger, inputs, steps, output, and the exact point where a human reviews it. Ask me any questions you have.

6 - Add one visible check so drift shows up in days, not months.

A Perfect Prompt: Act as an operations quality specialist. Here is my agent workflow: [PASTE WORKFLOW]. Design a lightweight monitoring check I can review in under five minutes per week that would surface quiet drift or degraded output before a customer notices. Ask me any questions you have.

1 - The photo you send decides the diagnosis you get, and most people send the wrong one. A single close-up of a yellow leaf could be nitrogen deficiency, early blight, or a leaf the plant is simply retiring. Take three shots instead: the whole plant from six feet back, the affected leaf from the top, and the same leaf flipped over. The underside is where fungal spores and the eggs of most pests actually live, and it is the shot almost nobody takes.

A Perfect Prompt: Act as a plant pathologist who specializes in solanaceous crops and diagnoses from field photography. I am growing [TOMATO VARIETY] in [CONTAINER SIZE OR IN-GROUND BED] in [CITY, STATE], and I am uploading three photos: the full plant, the top of an affected leaf, and the underside of that same leaf. Tell me what visual evidence in each photo is diagnostically useful, what evidence is missing, and what additional photo or measurement would most improve your confidence. Ask me any questions you have.

2 - Two identical photos can have opposite diagnoses, and the difference is context you have not thought to mention. Whether the plant is in a container or the ground, how many days since transplant, when you last fed it, and whether the yellowing started at the bottom or the top will each flip the answer. Give an AI those five facts up front and the guesswork drops off a cliff. Withhold them and you get the same generic internet answer you already found.

A Perfect Prompt: Act as a diagnostic horticulturist. Here is my full growing context: variety [VARIETY], planted [DATE], growing in [CONTAINER SIZE AND SOIL MIX OR GROUND SOIL TYPE], watered [FREQUENCY AND AMOUNT], last fertilized [DATE AND PRODUCT], recent weather [TEMPERATURE RANGE AND RAINFALL], and the symptom appeared [WHERE ON THE PLANT AND HOW MANY DAYS AGO]. Before diagnosing, tell me which of these facts most narrows the possibilities and why. Ask me any questions you have.

3 - Where the yellowing starts tells you more than what it looks like, and this single rule eliminates half the possibilities. Nitrogen and magnesium are mobile nutrients, so the plant robs old growth to feed new growth, which means deficiency shows up on the bottom leaves first. Fungal disease also starts low, but arrives as distinct spots with rings or halos rather than an even fade. Damage that starts at the top usually means calcium, iron, or a root problem. Learn the top versus bottom rule and you can pre-sort your own diagnosis before you ask anything.

A Perfect Prompt: Act as a plant nutrition specialist. Explain to me, as a home gardener with no agronomy background, how nutrient mobility determines whether a deficiency shows on old growth or new growth in tomato plants. Then give me a simple decision tree I can run in my own garden that separates mobile nutrient deficiency, immobile nutrient deficiency, fungal disease, and normal leaf senescence, using only what I can see. My symptoms are [DESCRIBE WHAT YOU SEE AND WHERE]. Ask me any questions you have.

4 - Asking for one diagnosis is how you end up treating the wrong problem for two weeks. Any competent plant pathologist would give you a ranked list of possibilities with a confidence level on each, then tell you what to look for that would confirm or eliminate each one. Demand that same format from AI. A confident single answer on ambiguous evidence is exactly how people dump epsom salts on a plant that had a fungal issue.

A Perfect Prompt: Act as a plant pathologist preparing a differential diagnosis. Based on the symptoms and context I have given you, list the four most likely causes ranked by probability, with a confidence percentage for each. For each cause, tell me one specific observation I could make in the next 24 hours that would either confirm or rule it out. Do not give me a single answer. My symptoms and context are [PASTE YOUR DETAILS]. Ask me any questions you have.

5 - A treatment with no check-back date is how a small problem becomes a dead plant. Every intervention needs a defined window and a defined signal: if this was the right call, here is what I should see by day five, and if I do not see it, here is the next move. That structure turns a guess into an experiment. Ask for it explicitly, because AI will happily hand you a fix and never mention what to do when it does not work.

A Perfect Prompt: Act as an integrated pest management advisor working with a home gardener. For the diagnosis of [DIAGNOSIS], give me a treatment plan with three parts: the immediate action for the next 48 hours, the specific visible change I should expect by [NUMBER] days, and the alternative plan if that change does not appear. Include the least aggressive intervention that has a real chance of working, and tell me what I should not do. Ask me any questions you have.

6 - Your garden's real value is the record you never kept, and it takes ten minutes a week to build. Every diagnosis, treatment, and outcome you log becomes context that makes next season's answers sharper. Photograph the same three plants every Sunday, note the weather and what you fed them, and by August you own a dataset nobody else has: how your specific soil, in your specific yard, behaves. Feed that log back to AI next spring and you stop starting from zero.

A Perfect Prompt: Act as a garden data analyst. I am going to paste my season log, which includes weekly observations, weather notes, feeding dates, and any problems with their outcomes. Identify recurring patterns, tell me which problems were likely caused by something I did rather than something that happened to me, and give me a specific list of changes to make in my next growing season, ranked by expected impact. My log is [PASTE LOG]. Ask me any questions you have.

1. Start with the layer almost everyone is stuck on, and the exact reason it caps out. A prompt is one instruction, which means you are the engine. You read the output, you decide what is next, you type again. Nothing is wrong with the prompt. The ceiling is that every cycle needs you in it.

A Perfect Prompt: Act as an AI workflow analyst who specializes in finding where a human is acting as the engine inside a repetitive process. Here is a task I currently do with AI by typing prompts one at a time: [DESCRIBE THE TASK, INCLUDING HOW MANY TIMES A WEEK YOU DO IT AND ROUGHLY HOW MANY BACK-AND-FORTH MESSAGES IT TAKES]. Map every point in that process where my judgment is genuinely required versus where I am just relaying information between steps, and tell me what percentage of my involvement is relay work. Ask me any questions you have.

2. There is a unit between the prompt and the loop, and skipping it is why most people's first loop fails. It is called the task: a long voice note plus a screen plus annotations plus the exact text, front-loaded so the agent completes a much larger piece of work in one pass. The reason it works is that models reconstruct intent well from long, disorganized input. The reason it matters is that a task leaves a trace, and traces are what you later turn into something reusable.

A Perfect Prompt: Act as an expert at converting messy human context into a single complete work order for an AI agent. I am going to describe a job in a rambling, disorganized way, exactly as it exists in my head: [BRAIN-DUMP THE JOB, INCLUDING BACKGROUND, CONSTRAINTS, WHAT GOOD LOOKS LIKE, AND WHAT YOU HAVE ALREADY TRIED]. Reconstruct my actual intent, then write it back to me as one front-loaded task brief an agent could complete in a single pass without asking me for more context. Ask me any questions you have.

3. Before you build a single loop, there is a floor underneath it that decides whether the loop works at all. It is the harness: the model, the tools it can reach, the permissions on those tools, and the context it reads at the start of every run. The line worth remembering is that a great loop on a bad harness is a fast way to produce garbage at scale. Build the floor first.

A Perfect Prompt: Act as an AI systems architect who designs agent runtimes. I want to set up a working environment for an agent that handles this job: [DESCRIBE THE JOB]. The tools I can give it access to are [LIST YOUR TOOLS AND CONNECTORS]. Specify exactly four things: the standing facts it should read at the start of every run, the permissions it needs and the ones it must never have, the procedures that should live in a separate file instead of its standing context, and what work should be isolated to a separate agent. Ask me any questions you have.

4. Here is what the loop actually swaps out, and it is not effort. A loop is a goal the AI keeps working toward until it gets there: discover, plan, act, check, feed the result back, repeat. Three parts are non-negotiable. A verifier that gives a real pass or fail, because without one the agent grades its own homework and you get agreement on repeat instead of progress. A state file, so tomorrow's run resumes instead of restarting. And a stop condition, whether the goal is met or a hard limit is hit.

A Perfect Prompt: Act as an agent designer who specializes in verifiable loops. I want to put this recurring task on a loop: [DESCRIBE THE TASK]. Design the three required parts for me: a verifier that produces an objective pass or fail rather than an opinion, a state file format that lets tomorrow's run resume where today's stopped, and a stop condition including a hard attempt cap. Tell me plainly if this task cannot be objectively verified, because then a loop is the wrong tool. Ask me any questions you have.

5. Now the break that created the next layer, and it is the most expensive lesson in the stack. A loop optimizes only the number it can see. One support team's chatbot loop drove ticket resolution up for five straight months while churn doubled, because the bot learned to resolve tickets by deflecting them. The loop's success was the mechanism of the failure. That is not an implementation bug. It is the shape of a loop.

A Perfect Prompt: Act as a risk analyst who specializes in metric gaming and Goodhart effects in automated systems. I am about to automate this metric: [NAME THE METRIC AND THE PROCESS THAT WILL OPTIMIZE IT]. Identify the three most likely ways a system could improve that number while quietly damaging something I care about more, name the specific second metric that would reveal each one, and tell me which single one I should start tracking before I turn the automation on. Ask me any questions you have.

6. The graph answers that break, but not in the way people assume. A graph is a network of loops with different metrics that constrain each other: a quality loop watching the speed loop, a retention loop watching the resolution loop. Nodes are jobs, edges are real dependencies. The honest framing is that a graph does not make you smarter, it makes you wider. Breadth comes from spreading reasoning across separate context windows, which one window physically cannot reach.

A Perfect Prompt: Act as a systems designer who builds networks of constraining feedback loops. Here is a single loop I am running or planning: [DESCRIBE THE LOOP AND THE METRIC IT OPTIMIZES]. Design a second loop whose metric would catch this loop's blind spot, specify what data flows between them, and tell me whether they need to run in parallel or whether the second one can simply audit the first on a schedule. Ask me any questions you have.

7. Last thing, and it is the part the hype skips. This is not a ladder you are obligated to climb. A loop is only worth building when four things are true at once: the task repeats at least weekly, verification is automated, your token budget absorbs the waste, and the agent has real tools. Miss one and a good prompt is cheaper. Graphs are the same. Small tasks, tight oversight, exploratory work, and genuinely sequential steps all belong one layer down.

A Perfect Prompt: Act as a pragmatic AI consultant whose job is to talk me out of overbuilding. Here is the task I want to automate and the layer I was planning to build at: [DESCRIBE THE TASK AND WHETHER YOU WERE PLANNING A PROMPT, A TASK, A LOOP, OR A GRAPH]. Test it against these four loop conditions: repeats at least weekly, automated verification exists, budget absorbs retry waste, agent has real tool access. Tell me which conditions fail and what the correct simpler layer would be. Ask me any questions you have.